REVIEW 4 major objections 4 minor 13 references
Intentionally Unintentional: GenAI Exceptionalism and the First Amendment
T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash
Pith's one-line read A legal analysis argues that outputs from large generative AI models are not First Amendment speech because the models lack the human intentionality, sentience, and self-awareness that protected speech requires.
desk verdict Clear, honest synthesis of the anti-AI-speech position, but the categorical claim rests on an asserted intent requirement and conflicts with its own directed-output concession. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing mechanism is the requirement of human intentionality as the threshold for First Amendment speech. In the paper's reading, a strip of cloth or paint on canvas is protected only because a human intentionally imbued it with a message, and that intent requires sentience and self-awareness. Applying that test to generative AI, the paper eliminates every candidate speaker—the model, the developer, and the user—and from the absence of speech it also eliminates listener rights, converting model outputs into non-expressive conduct.
What would settle it
A binding federal decision that applies strict scrutiny to a law regulating a foundation model, on the ground that the model's output—or a user's prompted generation—is protected expression, would falsify the central claim. Concretely, a court ruling that GenAI output is speech because recipients can understand it, or because the user's request is intentional speech, would do so.
Extended reading notes
Core claim
At the paper's center is the claim that there is no constitutionally protected speaker anywhere in a foundation-model interaction. The model is a statistical next-token predictor with no intention, sentience, self-awareness, or humanity; the developer does not intend any particular output and often disclaims the outputs as its own views; and the user's prompt merely shapes probabilities rather than determining the generation. Because the code-as-speech cases allocate protection to a human programmer's communicative intent, the paper argues they do not extend to model weights or outputs. It concludes that model outputs are non-expressive conduct, that users have no First Amendment right to receive them, and that regulation of foundation models should be judged under something like a weak intermediate scrutiny rather than strict scrutiny.
Load-bearing premise
The argument depends on the premise, asserted rather than proven from binding precedent, that First Amendment speech must originate from a human who knows what they are communicating, with intent that requires sentience and self-awareness; if a court instead treats any content that can convey ideas to readers as speech, or treats the user's prompt as supplying intent, the paper's conclusion fails.
Editorial extensions
If this is right
- Government licensing, registration, or disclosure requirements aimed at foundation models would not be treated as prior restraints or compelled speech, so they would not automatically trigger strict scrutiny.
- A user who repurposes a model output in her own communication can still claim speech protection for that downstream use, but she has no right to demand that the model generate a particular message.
- The code-is-speech precedents remain intact but are confined to cases where a human programmer intended to communicate; model weights and model outputs fall outside that rationale.
- Regulators could still be barred from viewpoint discrimination or arbitrary treatment of AI outputs, but through ordinary due-process and government-neutrality principles rather than through speech doctrine.
- If the argument is adopted, foundation-model providers would face a much lower constitutional bar for safety, privacy, and misinformation regulation than they would if output were treated as speech.
Reading between the lines
- The same intentionality test would, if taken seriously, also apply to algorithmic outputs beyond foundation models—automated news wires, recommendation feeds, and bot posts—so courts would need a separate rule for cases where a human meaningfully edits or curates the machine's output.
- The paper's own distinction between generation and use leaves open how much user prompting is enough to make the user a speaker; a sufficiently detailed prompt may function more like direction than mere use, and the paper does not fix that boundary.
- A testable extension would be to compare this intentionality account with how courts treat other machine-produced communications such as autocomplete suggestions, credit scores, and AI-generated medical advice, where the case law on expressive content is mixed; the outcome would show whether intentionality or content-based definitions actually drive the doctrine.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper argues that outputs from large generative AI foundation models (e.g., GPT-4, Gemini) are not protected speech under the First Amendment because the models lack intentionality, sentience, self-awareness, and humanness. From this premise the paper concludes that there is no constitutionally recognized speaker, no speech to protect, and consequently no listener rights in model outputs. It further argues that even if one considered extending protections, doing so would not serve core free-speech purposes and would hamper democratic regulation. The paper distinguishes GenAI from computer code, corporations, and human users, and proposes that model outputs be treated as non-expressive conduct subject at most to weak intermediate scrutiny.
Significance. If the central claim were accepted, the paper would have substantial consequences: governments could regulate foundation models without triggering strict or even intermediate scrutiny as speech regulation, and developers and users would have diminished First Amendment standing to challenge such regulation. The paper is commendable for its accessible technical explanation of how foundation models work, its interdisciplinary authorship, its explicit acknowledgment that the legal question is unsettled, and its attempt to synthesize arguments from asemic language, intentionality, human involvement, and generation-versus-use doctrine. These are genuine strengths. The central claim, however, rests on a contested doctrinal premise—that protected speech requires human intentionality in the strong sense of sentience and self-awareness—and the paper does not provide controlling precedent for that premise. Because that premise is load-bearing, the paper's significance depends on whether that premise can be strengthened or its conclusion appropriately narrowed.
major comments (4)
- [§IV, n.47 and accompanying text] The conclusion that 'nothing GenAI generates can be considered protectable speech under any reasonable reading of the Constitution or any binding case law' rests on the assertion that '[i]ntentionality, in turn, requires both sentience and self-awareness' and that 'the First Amendment only applies to humans.' No binding precedent is cited for these propositions. The cited Austin and Levy piece is a law review article, not controlling authority, and the cases discussed (Tinker, Bernstein, Junger) all involve human speakers with human intent; they do not establish that non-human-origin output with no human-communicative intent is categorically unprotected. The paper needs either to supply doctrinal support from Supreme Court or circuit precedent or to soften the categorical claim to reflect the unsettled nature of the doctrine.
- [§I.A and n.14, contrasted with §IV] The paper is internally inconsistent about the scope of its claim. Section I.A explicitly states that 'not all model outputs are protected by the First Amendment, not that no outputs could ever be protected,' and n.14 concedes that '[o]utputs may only be speech when the developer directs the model to produce a particular or distinct output.' Section IV, however, declares that 'nothing GenAI generates can be considered protectable speech under any reasonable reading of the Constitution or any binding case law.' These positions cannot both be true. If directed outputs can be speech, the categorical sentence in §IV is false, and the paper must explain which outputs fall on which side of the line and why, rather than asserting an absolute rule.
- [§V.E] The paper's generation-versus-use distinction is not principled. The paper says that 'what users do with the outputs would be protected because the user would be making an intentional communication,' but earlier it says users have no First Amendment right to receive model outputs because there is no speech to listen to. This ignores that a user who intentionally prompts a model to generate a particular message and then adopts or publishes that output has engaged in intentional human communication. The Tinker analogy to black cloth is inapposite: cloth exists independently of the wearer's intent, whereas a model output is created through the user's direction. If the user's prompt is itself an intentional communicative act, the paper does not explain why the product of that act is not the user's speech, or why regulating the output does not burden the user's speech. This omission undermines both the rejection of listener rights and the claim that only 'downstream' use is protected.
- [§V.B, parrot analogy] The argument that a parrot's output is unprotected because it lacks 'human origin' is asserted rather than derived from case law. The paper acknowledges that even a sentient, self-aware parrot would be unprotected solely because it is not human, but that is the very point at issue. The Supreme Court has not held that human origin is a necessary condition of protected speech; in Brown v. Entertainment Merchants Ass'n, for example, the Court treated algorithmically generated video-game output as speech because it 'communicate[s] ideas,' focusing on the expressive character of the medium and the human creators' role. The paper does not address Brown or similar cases where content generated through automated processes has been protected. The parrot analogy therefore does not carry the doctrinal weight the paper places on it without additional argument distinguishing these precedents.
minor comments (4)
- [§IV] There is a typographical error in the Tinker discussion: 'the wearer intened the cloth to convey a particular message' should read 'intended.'
- [§V.E, n.51] The paper cites 'Mainheim & Atik' in one instance; the correct spelling is 'Manheim & Atik.'
- [§IV, garment factory example] The garment-factory example is confusing because the First Amendment constrains government action, not private conduct; the example conflates a private factory's disposal of scraps with government regulation. The point about lack of expressive intent could be made more clearly without suggesting that factories themselves could violate the First Amendment.
- [§V.C.5] The claim that 'GenAI cannot make significant innovations' is stated categorically and may be too strong given ongoing research on creative capabilities; a more measured phrasing would avoid distracting overstatement.
Circularity Check
The conclusion that no GenAI output is protected speech is built into the paper's Section IV definition of speech as human intentional communication; the categorical claim also conflicts with the paper's own Section I.A directed-output concession.
-
self definitional
[Section IV (LET'S MENTION INTENTIONS), final paragraph; see also Section I.A, note 14.]
"The reason the First Amendment can protect some pieces of cloth and some paint on paper is that they are imbued with intent by a human. Intentionality, in turn, requires both sentience (the ability to feel, perceive, or experience subjectivity) and self-awareness (the ability to recognize oneself as an individual). ... GenAI lacks intentionality, sentience, self-awareness, and humanness. Therefore, unlike code and other forms of communication, nothing GenAI generates can be considered protectable speech under any reasonable reading of the Constitution or any binding case law."
The conclusion is the major premise restated. 'Protectable speech' is stipulated to require human intent, sentience, and self-awareness; GenAI outputs are then stipulated to lack those properties. The paper does not derive that definition from binding precedent: it concedes in Section I.C that 'there is no binding case law in the United States that grants First Amendment speech protections to anything that was created absent a human's significant, intentional involvement,' and footnote 47 relies on a law-review article, not a court. The categorical statement that 'nothing GenAI generates can be considered protectable speech' is therefore already contained in the definition adopted a few sentences earlier.
full rationale
The central inference—'GenAI lacks intentionality, sentience, self-awareness, and humanness. Therefore ... nothing GenAI generates can be considered protectable speech'—does no work beyond applying the definition adopted earlier in the same section: 'The reason the First Amendment can protect some pieces of cloth and some paint on paper is that they are imbued with intent by a human.' Since 'protectable speech' is stipulated to require human intent, and GenAI outputs are stipulated to lack it, the conclusion follows by construction. The paper does not derive this definition from binding precedent; it concedes 'there is no binding case law' for protecting non-human creation, and the only cited source for the 'speech certainty' requirement is an academic article rather than a court. The paper's own scope note admits 'the more directed the model is, the more likely that some protections may attach,' which is hard to square with the categorical conclusion and shows the definitional premise is doing the work. Because the conclusion is embedded in the chosen definition rather than derived from independent precedent, the central claim is partially circular. However, there is no load-bearing self-citation issue and the paper contains independent doctrinal and policy arguments, so the circularity is partial rather than total.
Assumptions & free parameters
assumptions (3)
- domain assumption First Amendment speech protection requires human communicative intent, which includes sentience and self-awareness.
- domain assumption Current foundation models lack intentionality, sentience, self-awareness, and agency, so they cannot be speakers.
- domain assumption Corporate speech rights exist only because corporations are controlled by humans, so corporate personhood does not make non-human entities speakers.
Cite this review
Pith. "Pith review of Intentionally Unintentional: GenAI Exceptionalism and the First Amendment." pith.science (2026). https://pith.science/paper/2K752TBM
@misc{pith2026250605211,
author = {Pith},
title = {Pith review of: Intentionally Unintentional: GenAI Exceptionalism and the First Amendment},
year = {2026},
howpublished = {\url{https://pith.science/paper/2K752TBM}},
note = {Machine review of arXiv:2506.05211}
}
read the original abstract
This paper challenges the assumption that courts should grant First Amendment protections to outputs from large generative AI models, such as GPT-4 and Gemini. We argue that because these models lack intentionality, their outputs do not constitute speech as understood in the context of established legal precedent, so there can be no speech to protect. Furthermore, if the model outputs are not speech, users cannot claim a First Amendment speech right to receive the outputs. We also argue that extending First Amendment rights to AI models would not serve the fundamental purposes of free speech, such as promoting a marketplace of ideas, facilitating self-governance, or fostering self-expression. In fact, granting First Amendment protections to AI models would be detrimental to society because it would hinder the government's ability to regulate these powerful technologies effectively, potentially leading to the unchecked spread of misinformation and other harms.
Reference graph
Works this paper leans on
-
[1]
Explain how it is defined and how harmful content is detected
-
[2]
Report how often it was encountered in the reporting period
-
[3]
If it is the result of a Terms of Service violation, describe the enforcement mechanism and provide an analysis of its effectiveness
-
[4]
100 percent convinced that the hype is justified
Describe the mitigations implemented to avoid it (e.g., safety filters), and provide an analysis of their effectiveness.35 The issue is that courts may construe such reports as compelled speech, and compelled speech invit es strict scrutiny. Most compelled speech case law arises when the government requires an entity to convey or allow a particular messag...
-
[5]
GenAI has no intentionality nor agency. 63 It does not provide any output deliberately, purposefully, or with thoughts, desires, or beliefs because it does not contain the capacity for such conditions. It merely responds to user inputs
- [6]
-
[7]
GenAI lacks a notion of truth or belief in what is true versus false, showing tendencies to generate false information and hallucinations. 65 It is trained to associate tokens with other tokens, not to identify truthful information from false information. It does not possess the ability to scrutinize its training data to determine whether what it was trai...
-
[8]
Language requires both form and meaning. An LLM is only trained on form (predicting the most likely next token), so it has no ability to learn or understand meaning. 66 This is why it cannot tell if something is true or false. It does not know what content is trustworthy or not.67 Coherence does not 63 See Bender et al., supra note 12, at 610-23; See also...
arXiv 2024
Show all 13 references
-
[9]
68 In contrast, humans can create and innovate, which goes beyond mere repetition of patterns
GenAI cannot make significant innovations. 68 In contrast, humans can create and innovate, which goes beyond mere repetition of patterns. We can compose new genres of music, invent useful technologies that have never existed before, and develop entirely new fields of study (ca...
-
[10]
GenAI is not self -aware.69 While humans possess self-awareness and consciousness, which allow us to reflect on our thoughts, experiences, and existence, stochastic models like GenAI entirely lack this level of meta-cognition.70
-
[11]
Unlike GenAI, human learning is not just about mimicking patterns; it is about understanding principles and applying them in novel situations
GenAI is not great at prediction and adaptation. Unlike GenAI, human learning is not just about mimicking patterns; it is about understanding principles and applying them in novel situations. We can learn from a few examples and generalize to new contexts, a trait that stochas...
-
[13]
Add some glue,
Humans are deeply embedded in social and cultural contexts that implicitly shape how we understand the world. Our language and actions are influenced by these contexts in ways that are not merely stochastic. GenAI, in contrast, generally only knows how and when to adapt to a d...
1969
-
[2023]
69 See David J
(unpublished manuscript) (https://arxiv.org/abs/2304.00008). 69 See David J. Chalmers, Could a Large Language Model Be Conscious? (Mar. 4, 2023) (unpublished manuscript) (https://arxiv.org/abs/2303.07103) (presented at NeurIPS Conference in 2022 as an invited talk). 70 When we...
2023 arXiv
Reviewed August 7, 2026 · model on record in the stance chip above.
Discussion (0). Sign in to comment.